Customer Service & Support
Routine contacts are increasingly resolved end to end by software. What reaches a person is what the software could not finish. The honest question is not whether that is happening — it is what it does to the work that is left, and that turns out to be less settled than either side claims.
The routine queue is being automated away; what survives sits upstream of it — deciding what the agent handles, what it draws on, and where it hands over.
Measured
| Occupation | Measured exposure | Observed automation share |
|---|---|---|
| Customer service representatives | 100th percentile | 34% |
| Computer support specialists | 95th percentile | 73% |
The automation share is measured from observed AI usage, published by Anthropic under CC BY 4.0 — a vendor reporting on its own product, and worth reading as such. It describes how people use AI for this work, not how much of the work AI can do. Exposure percentiles are our composite of the measured sources set out in the methodology.
The shift
The occupation is projected to contract over the decade, and the stated cause is that simple tasks stop reaching a person at all. Note the second number though: this is a contraction, not a disappearance.
Customer service representatives: 2,666,000 jobs, −5% projected 2025–35, a fall of about 141,800 — against 289,500 openings a year, almost all replacing leavers. Median pay $44,770.
US Bureau of Labor Statistics, Occupational Outlook Handbook · 2026-08-27 · verified 2026-09-07
The reason given is not that the work became less valuable. It is that the easy half of it stopped being routed to a human in the first place.
“Self-service systems, social media, and mobile applications enable customers to do simple tasks” without interacting with a representative.
US Bureau of Labor Statistics, Occupational Outlook Handbook · 2026-08-27 · verified 2026-09-07
The technical-support half of this cluster is projected to shrink on the same mechanism, and here the statistical agency names the technology directly rather than describing it.
Computer user support specialists: 750,600 jobs, −3% projected 2025–35, “as organizations continue to implement automated tools, such as chatbots, for troubleshooting”. Network support, a different job, is projected to grow 1%.
US Bureau of Labor Statistics, Occupational Outlook Handbook · 2026-08-27 · verified 2026-09-07
The queue
The clearest official statement of what happens to the work that remains, and it is carefully hedged by the people who wrote it. Both halves of the sentence matter: more complex work, and fewer people doing it.
Automation “may free up some computer user support specialists to handle more complex cases” — “but fewer are expected to be needed overall”.
This is the agency's stated reasoning for a projection, not an observed measurement of anyone's workload.
US Bureau of Labor Statistics, Occupational Outlook Handbook · 2026-08-27 · verified 2026-09-07
The one randomised field experiment on this question found that human rescue works for technical escalations and works less well for emotional ones — and that the people handling them disengaged. The cause it identifies is escalation design and timing, which is a management decision, not a personal failing.
Human intervention preserves quality in technical escalations but “is less effective in algorithm-triggered emotional escalations”, where “workers showed lower engagement: they sent fewer messages”.
One platform, one market. The abstract reports directions rather than effect sizes, and the same paper contains a finding that cuts the other way — it is in the next section rather than left out.
Wang, Zhu, Feng, Lu and Jia, field experiment at Alibaba, arXiv:2605.14830 · 2026-05-14 · verified 2026-09-07
Deploying the agent made conversations shorter and, on the contacts it handled, worse. That combination is the argument for someone owning what the agent is allowed to take.
AI deployment “reduces average chat duration”, has limited effect on retrials, and “substantially lowers ratings for AI-eligible chats”.
Wang, Zhu, Feng, Lu and Jia, Alibaba field experiment · 2026-05-14 · verified 2026-09-07
A named analyst describes the same effect as a structural consequence of a design choice — the easy questions siphoned off, leaving a permanent escalation role — rather than as something the people doing it should get better at absorbing.
“smarter customer service tech, but more cognitive load on the people behind the screens” — Riccardo Pasto, Principal Analyst, Forrester
Said to a trade publication. There is no Forrester report behind it that we could locate, so it is an analyst's stated view rather than published research.
Riccardo Pasto, Principal Analyst, Forrester, in No Jitter · 2026-03-13 · verified 2026-09-07
Against this reading
The same experiment we cited above found the opposite of overload on the hard contacts: freed from the routine queue, workers gave the difficult conversations more attention, not less. This is the single strongest piece of evidence against the argument on this page.
The authors document “a positive spillover effect on AI-ineligible chats”, as treated workers adapted their workflow to devote greater attention to them.
Wang, Zhu, Feng, Lu and Jia, Alibaba field experiment · 2026-05-14 · verified 2026-09-07
A named research house predicts that half the companies who attributed support cuts to AI will reverse them — and its survey found that only a fifth of leaders had actually cut agent headcount for that reason at all.
By 2027, half of companies attributing headcount reduction to AI “will rehire staff to perform similar functions, but under different job titles”. Of 321 leaders surveyed, about 20% had reduced agent headcount because of AI.
A prediction, not a measurement, and cited via TechRepublic because Gartner's own page is not publicly fetchable. The 20% figure is the load-bearing one and it undercuts every hard-automation headline in this space.
Gartner, survey of 321 customer service and support leaders · 2026-02-03 · verified 2026-09-07
The same analyst house disputes the attribution the whole page rests on: that the cuts were caused by AI rather than by the economy.
“Most recent workforce reductions were influenced by broader economic conditions rather than automation alone.” — Kathy Ross, Senior Director Analyst, Gartner
Kathy Ross, Senior Director Analyst, Gartner, via TechRepublic · 2026-03-02 · verified 2026-09-07
Consumers are moving against automated support, not toward it — which is a commercial constraint on how far deflection can actually go, whatever the technology can do.
79% strongly prefer interacting with a human, 84% believe human agents are more accurate, and 81% believe AI is used primarily to save the company money.
SurveyMonkey, 2,017 US adults, fielded 10–11 December 2025, ±2.5pp · 2026-02-19 · verified 2026-09-07
The best-known study of AI in support measured the opposite of burnout. It looked at AI assisting people rather than replacing the routine queue, which is a different question — but it is the one most often cited in this debate, and it does not say what it is used to say.
Across 5,179 support agents, AI assistance raised issues resolved per hour by 14% on average and 34% for novices, improved customer sentiment, and increased employee retention.
Fieldwork ran in 2020–21, before agentic deflection existed. It measures a copilot beside a human, not an agent in front of one.
Brynjolfsson, Li and Raymond, Quarterly Journal of Economics · 2025 · verified 2026-09-07
On the record
The opening claim, made jointly with the company's own AI supplier. It set the terms of the entire debate that followed and it was, from the beginning, a company describing its own deployment.
An AI assistant handling two-thirds of service chats in month one — 2.3 million conversations, “the work of 700 full-time agents” — with resolution times down from 11 minutes to under 2.
Dated 2024, self-reported, and published alongside the vendor supplying the model. Not an independent measurement of anything.
Klarna, company press release, with OpenAI · 2024-02-27 · verified 2026-09-07
Fifteen months later the same chief executive said the cost-led version of this had produced worse service, and the company began recruiting human agents again.
Cost had been “a too predominant evaluation factor”, and “what you end up having is lower quality”. — Sebastian Siemiatkowski, CEO, Klarna
The Bloomberg interview is the primary source; this is trade press reporting it. We have not read the Bloomberg original.
Sebastian Siemiatkowski, CEO, Klarna, to Bloomberg, reported by CX Dive · 2025-05-08 · verified 2026-09-07
The reversal was not a retreat from automation. It was a repricing of human contact as a premium tier — which is a different and more consequential claim about this job than either the boosters or the sceptics usually make.
“offering human customer service is always going to be a VIP thing.” — Sebastian Siemiatkowski, on stage at SXSW London
Sebastian Siemiatkowski, CEO, Klarna, reported by TechCrunch · 2025-06-04 · verified 2026-09-07
Eighteen months after the reversal the company raised its automation claim again — while its actual cost of running customer service went up, not down. Both numbers were disclosed on the same earnings call.
The AI agent described as doing the work of 853 full-time staff and saving $60M, in a quarter where customer service and operations costs rose to $50M from $42M a year earlier.
Klarna Q3 2025 earnings call, reported by CX Dive · 2025-11-20 · verified 2026-09-07
Where it actually landed is the part worth reading closely, because it is not an upskilling story. The hard residual is being staffed gig-style, recruited from the customer base, while headcount keeps falling.
“it means it's going to be the cheap customer service” — Siemiatkowski on AI-handled support, describing an Uber-style model for the human tier.
This is the outcome to argue with rather than the one to celebrate. Concentrating the difficult work does not automatically raise what it pays; at Klarna it accompanied a weaker employment relationship.
Sebastian Siemiatkowski, CEO, Klarna, on the 20VC podcast, reported by CX Dive · 2026-02-20 · verified 2026-09-07
The second clean public example. Note where the cut fell and where the people went — this is routinely miscited as a statement about sales headcount, and it is not.
“I've reduced it from 9,000 heads to about 5,000, because I need less heads.” — Marc Benioff, on support, with staff redeployed into sales and services.
Fortune's rendering of a podcast remark. We have not timestamped it against the audio, and no Salesforce filing states this figure.
Marc Benioff, CEO, Salesforce, on The Logan Bartlett Show, reported by Fortune · 2025-09-02 · verified 2026-09-07
Analysts
The analyst view of what the machine still cannot do is narrower and more specific than the marketing on either side, and it names the three things the residual queue is made of.
“AI simply isn't mature enough to fully replace the expertise, empathy, and judgment” human agents provide. — Emily Potosky, Senior Director Research, Gartner
Emily Potosky, Senior Director Research, Gartner, via CX Network · 2025-09-15 · verified 2026-09-07
The most useful framing anyone has put on this publicly: the load did not appear because the work got harder to do. It appeared because of how the handover was designed.
Siphoning off the easy questions pushed agents into a perpetual escalation role — “more cognitive load on the people behind the screens”.
An analyst's view given to a trade publication, not a published study. It is the mechanism this page argues for, and it is not measured.
Riccardo Pasto, Principal Analyst, Forrester, in No Jitter · 2026-03-13 · verified 2026-09-07
Where the work goes
The destination exists as a real, named, posted job at a real company, and what it asks for is precisely the upstream work: deciding what the agent handles, keeping what it draws on correct, and designing the handover.
Notion, AI Conversation Designer, Customer Support: $136,000–$155,000, owning “intent architecture end-to-end: taxonomy design, hierarchical intents, coverage mapping, and gap analysis”.
The listing closed in July 2026; the page still carries the band and the responsibilities. And here is what would be dishonest to leave out: this is one posting at one technology company. No statistical agency or labour-market dataset we could find sizes this role as a category. It is a narrow, well-paid door, not somewhere an occupation of two and a half million people can walk through.
Notion job posting, via Built In NYC · 2026-07 · verified 2026-09-07
The move is being made by real people with checkable names, and it goes in the direction this page describes: away from answering the repetitive contacts, toward owning what the system answers them with.
At RB2B, a head of technical operations became head of AI, moving “from repetitive support questions to managing knowledge and improving the system behind it”.
Published by a support-software vendor with an interest in this story, and the wording is Intercom's rather than Clarke's own. The people and job titles are named and independently checkable, which is why it is here; the survey figures in the same post are a vendor survey and are not.
Declan Ivory, VP of Customer Support, Intercom, describing Robb Clarke's move at RB2B · 2026-02-27 · verified 2026-09-07
Designs and owns the support stack around an AI agent - dialogue flows, deflection tuning, the answer base and escalation paths - accountable for answer quality and CSAT against the bot.
Read those skill lists honestly against your own. Someone doing this work today overlaps on the domain knowledge and generally not on the agent tooling, which places them adjacent to the destination rather than already in it. That gap is the work, and naming it is more useful than a score that flatters.
If this is your title
The measurements above describe occupations. They do not describe you. The audit reads your own evidence — what you have owned, and what of it is repeatable somewhere else.
Get Started